Bug 50516 – Avoid strlen() in AtomicString::fromUTF8

• Component: Web Template Framework
• Status: RESOLVED FIXED

[Patrick R. Gansterer]

see patch

[Patrick R. Gansterer]

Created attachment 75606 [details]
Patch

[Patrick R. Gansterer]

about 2% performance improvement

[Build Bot]

Attachment 75606 [details] did not build on win:
Build output: http://queues.webkit.org/results/6732040

[Eric Seidel (no email)]

Comment on attachment 75606 [details]
Patch

Woh.  Nearly a page of code for "2%"?  On what benchmark?  w/o more explanation, this is clearly not worth it.

[Darin Adler]

(In reply to comment #2)
> about 2% performance improvement

2% improvement by what measure on what platform?

I don’t agree with Eric in that I do think this change is probably a good idea. We can probably factor things so that a lot more of the code is shared and still retain the speed.

But I’d like to understand more about what is made faster by this.

[Patrick R. Gansterer]

(In reply to comment #5)
> (In reply to comment #2)
> > about 2% performance improvement
> 
> 2% improvement by what measure on what platform?

I did a javascript loop with (new DOMParser()).parseFromString("...", "text/xml") and did the same time measuremet like in WebCore/benchmarks/parser on my MacBook Pro.
I'll upload a test.

[Eric Seidel (no email)]

I think Darin and I actually agree. :)

[Patrick R. Gansterer]

Created attachment 76745 [details]
XML parser benchmark

[Patrick R. Gansterer]

I've done the benchmark now again and got over 3% only with this patch. Tested on a MacBook Pro 2.4 GHz, 4GB RAM, OS X 10.6.5.

                     avg        median    stdev
 
OS X 10.6 original   >4000      >4000     ;-)
r74063               2185.72    2182      12.126895728091345
r74063 with patch    2108.01    2095.5    99.8814792641759
                    -3.56%      -3,96%

[Darin Adler]

I feel conflicted about this. I like the idea of having a fast code path.

But we’ve long believed main cause of slowness with XML parsing is the approach of using libxml2. A serious effort to speed up XML parsing would probably involve switching to a new parser or creating our own.

Since one of the main issues with this patch is a bit too much replicated code, we should refactor this so we can share the code. We can possibly make the two fromUT8 functions share more code by using an inline function for everything after the call to calculateStringHashFromUTF8 or use a function template. Similarly, we can make calculateStringHashFromUTF8 share almost all the code by using a function template that is inlined.

I think if we do that we can land this, but maybe we shouldn’t land it as is.

It’s good to know it’s a measurable speedup on an XML parsing micro-benchmark.

[Eric Seidel (no email)]

My concerns are with maintainability of this code.  The benchmark needs to be checked in, so that 5 years form now, others can validate that this is still a hot spot.  The code needs to be well factored so that it can be re-used, and possibly removed at some future time when this is not a hot spot.

We've certainly added and removed optimizations over the years.  Each page of code we add to webkit has a maintenance cost.  I'm hoping to minimize that cost. :)

A recent example was the HTML parser.  Which had all sorts of crazy optimizations which we dropped on the HTML5 parser re-write.  Because testing (and finally splitting out the code into classes) showed they didn't matter.  If what Darin says is true (that this isn't really the important hotspot for XML, rather libxml is), if a big attempt to optimize XML really replaced our libxml implementation, this code might not be needed at all (and thus simply blulk).

But you've seen the shark profiles and I havent. :)

[Patrick R. Gansterer]

Created attachment 79107 [details]
Patch

[Early Warning System Bot]

Attachment 79107 [details] did not build on qt:
Build output: http://queues.webkit.org/results/7529117

[Patrick R. Gansterer]

Created attachment 79110 [details]
Patch

[Patrick R. Gansterer]

@darin: ping?

[Darin Adler]

Comment on attachment 79110 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=79110&action=review

> Source/JavaScriptCore/wtf/unicode/UTF8.cpp:334
> +            if (dataLength)
> +                *dataLength += 1;

Seems better to not have if statements like this inside relatively hot loops. We should instead have callers supply a place to put the length even if they won’t use it.

> Source/JavaScriptCore/wtf/unicode/UTF8.cpp:369
> +            if (utf16Length)
> +                *utf16Length += 2;

No reason to have these branches in since utf16Length is never null.

> Source/JavaScriptCore/wtf/unicode/UTF8.h:75
>      unsigned calculateStringHashFromUTF8(const char* data, const char* dataEnd, unsigned& utf16Length);
>  
> +    unsigned calculateStringHashAndLengthFromUTF8(const char* data, unsigned& dataLength, unsigned& utf16Length);

I’d group these together rather than leaving a blank line.

[Patrick R. Gansterer]

Committed r77297: <http://trac.webkit.org/changeset/77297>

[Patrick R. Gansterer]

Committed r77491: <http://trac.webkit.org/changeset/77491>

[WebKit Review Bot]

http://trac.webkit.org/changeset/77491 might have broken Leopard Intel Release (Build)